DEV Community

AI-KSK
AI-KSK

Posted on

Ideogram 4 With the KJ Prompt Builder: Composition as Data, Not Luck

#comfyui #ai #tutorial

Ideogram 4's differentiator is that it consumes JSON prompts: break the image into background, subject, text, colours, regions and spatial relations, then let the model build to that layout. The official ComfyUI flow makes you compute the JSON with an LLM first; the KJ plugin turns it into a visual composition console. This episode ships both - the official flow and the KJ director flow.

Evidence note: two workflow JSONs ship with this episode; model files, scheduler and guider nodes below are read from them. Strengths, weaknesses and the licensing caveat are the author's, as reported in the video.

When to use it

  • You need text rendered inside an image, posters or multi-region layouts.
  • You are tired of rerolling for composition and want to place elements.
  • You accept a model with built-in censoring; it is not a portrait model.

What is different in this workflow

  • JSON prompt instead of prose: the KJ node exposes width/height, a global description, background, style (none / photo / art with a substyle), aesthetic controls, a global palette of up to 16 colours, and per-region boxes.
  • Region boxes are typed: object or text, each with its own description and up to 5 colours.
  • Two outputs: prompt (goes to the text encoder) and import JSON (paste a JSON built elsewhere, e.g. by Gemini 3.1).
  • Conditional + unconditional dual model: prompt goes to the conditional UNET, an empty prompt to the unconditional UNET, and CFG scales the difference - the author explains this in the video.
  • Official flow is three fields: size, quality, prompt; use it to sample the model cheaply.

Step by step

  1. Start with the official flow: set ~2 MP, keep Quality, write a short prompt, run. The author says three fields are enough to judge the model.
  2. Move to the KJ flow. Set width/height (they define the canvas aspect).
  3. Write the global description - what the whole image is, not a fragment list.
  4. Add the background description; pick style: none, photo, or art plus a substyle.
  5. Set aesthetic direction (lighting, medium such as high resolution) and a global palette of up to 16 colours.
  6. Add region boxes: choose object or text, describe it, give it up to 5 colours.
  7. Or paste a JSON from an external LLM into the import field.
  8. Keep Chinese short - the model handles Chinese, but the author advises using it sparingly.
  9. Run; iterate by editing boxes rather than rerolling.

Parameter basis

  • Official flow (Ideogram 4 出图官流.json, 56 nodes): UNETLoader -> ideogram4_fp8_scaled.safetensors and ideogram4_unconditional_fp8_scaled.safetensors; CLIPLoader -> qwen3vl_8b_fp8_scaled.safetensors (type ideogram4); VAELoader -> flux2-vae.safetensors; Ideogram4Scheduler; KSamplerSelect -> euler; DualModelGuider; CFGOverride; ResolutionSelector; EmptyFlux2LatentImage; ConditioningZeroOut for the unconditional branch.
  • KJ director flow (Ideogram 4 + KJ Prompt Builder可视化构图导演.json, 43 nodes): same model stack plus Ideogram4PromptBuilderKJ, JsonExtractString, ComfyMathExpression, ComfyNumberConvert, PrimitiveInt and StringReplace nodes that turn the console state into a JSON prompt.
  • Palette limits: 16 colours globally, 5 per element.
  • Two model builds: FP8 and NVFP4 are both published; the author explains the CFG mechanism with the conditional/unconditional pair.

Limits

  • It is not a portrait model: the author is blunt that photorealism and pretty faces are not its strength, and it has built-in moderation.
  • Region control can hurt the hero: in the two-character wuxia test both characters came out unclear, swallowed by the mechanism, lighting symbols and architecture; in the alien standoff the protagonist was not prominent enough. The author's read is that forcing the bbox makes the model drop other things.
  • Small requested elements come out small: text and the lone adventurer were rendered smaller than intended, though in the right place.
  • Style can drift to cinematic even when a realistic image was asked for (the mountain rescue shot).
  • Licensing: the model is not fully open and commercial use is restricted; the author expects community finetunes to appear anyway.

Episode resources

Top comments (0)